| With the rapid development of Internet technology,more and more people are willing to share their opinions on social media.Focusing on a controversial topic,the discussions between users form argumentation.The purpose of argumentation mining research on argumentation text is to automatically extract claims and premises as well as their logical relationships from unstructured argumentation texts.Argumentation mining in social media text automatically discovers user claims and understand the reasons why users hold their own claims.Therefore,it has important academic and commercial value.The existing argumentation mining methods may be camped into feature-engineering machine learning based approach and deep learning based approach.The feature-engineering machine learning based approach mainly uses artificial constructed and selected features to represent claims or premises,which are often cumbersome and also difficult to model argumentation structures well.The deep learning based approach automatically learns the representations of the deep semantic features of argumentation texts.However,most proposed methods analyze the subtasks of argumentation mining individually while the internal associations between subtasks are ignored.To address these issues,this study first investigates the method for effectively identifying argumentation components through introducing the argumentation indicators to hierarchical attention neural network.On this basis,the joint learning method for argumentation component identification and relation extraction is studied.In this study,the argumentation component identification method which incorporates argumentation indicator information and hierarchical attention neural network is firstly investigated.Through introducing argumentation indicator information to hierarchical attention neural network,we design and implement a hierarchical attention neural network model with argumentation indicator.The hierarchical attention neural network effectively models the discourse information,and the introduction of argumentation indicator helps the model capture the context associations between the argumentation components.They are helpful to improve the performance of argumentation component identification task.The experimental results on two public argumentation mining datasets,namely Persuasive Essays(PE)and Microtexts(MT),show that the performance of hierarchical attention neural network with argumentation indicator improved the performance obviously,compared to many baseline methods.The F1 values for argumentation component identification on the two datasets are increased by 1.4% and 0.8%,respectively.Considering the strong associations between the argumentation component identification task and the argumentation relation extraction task,this study further investigates the joint analysis method for argumentation component identification and relation extraction.A multi-task learning framework with adaptive weighing losses strategy,which is based on multi-level sharing,is proposed.This framework firstly models the input documents by using a hierarchical attention neural network with argumentation indicator information,and then jointly learns the argumentation component identification task and the argumentation relation extraction task following multi-level shared mode.Finally,an adaptive weighing losses strategy based on homoscedastic uncertainty is used to dynamically adjust the loss function weights for each subtask during the training process.Experimental results on PE and MT show that the proposed multi-task learning framework improves the overall performance of the argumentation component identification task and the argumentation relation extraction task.The macro F1 values of two tasks on the two datasets are improved by 1.3% and 1.8%,respectively. |
